TLDR: This research introduces a multi-level framework to diagnose and evaluate the robustness of Large Language Models (LLMs) in tabular data feature engineering. It focuses on how LLMs identify key variables, understand relationships, and set decision boundaries. The study found that LLM robustness varies significantly across datasets, and that providing high-quality examples is crucial for improving performance. By using this framework, high-quality LLM-generated features can enhance prediction performance by up to 10.52%, offering a path to more reliable LLM-driven data science.
Large Language Models (LLMs) have shown great potential in various data science tasks, including feature engineering for tabular data. Feature engineering is a crucial step where raw data is transformed into features that better represent the underlying problem to predictive models. However, a significant challenge with LLMs is the variability and inconsistency in the features they generate, raising concerns about their overall reliability.
Researchers Yebin Lim and Susik Yoon from Korea University have introduced a novel framework designed to diagnose and evaluate the robustness of LLMs in this critical area. Their work, titled “Multi-level Diagnosis and Evaluation for Robust Tabular Feature Engineering with Large Language Models,” addresses how consistently LLMs perform feature engineering across different datasets and input conditions. You can find more details about their research here: Research Paper.
Understanding the Multi-level Framework
The core of their framework is a multi-level approach that mirrors how human domain experts approach feature engineering. It focuses on three key aspects:
- Level 1: Identifying Key Variables (Golden Variable): This level assesses whether LLMs can correctly pinpoint the most important variables that are highly relevant to predicting a target outcome. For instance, in a diabetes prediction task, ‘Glucose’ would be a key variable. The framework tests this by introducing variations in variable descriptions and examples to see if the LLM consistently ranks the correct variables as important.
- Level 2: Understanding Variable-Class Relationships (Golden Relation): Here, the framework checks if LLMs can accurately determine the relationship between key variables and the target classes. For example, understanding that high glucose levels are positively linked to diabetes. Robustness is tested by altering sample quality and mixing variable values.
- Level 3: Setting Decision Boundaries (Golden Value): This level evaluates the LLM’s ability to identify specific values that effectively separate different classes. An expert might know that a glucose threshold above 100 indicates diabetes. The framework examines if LLMs can consistently provide stable decision boundaries under different input conditions.
Diagnosing Reliability
The diagnosis part of the framework involves introducing various changes to the input provided to the LLM, such as different variable descriptions, ordering, sample quality, and sampling methods. By observing how the LLM’s responses change, the researchers can categorize outputs into ‘high-score’ cases (where responses align with domain knowledge) and ‘low-score’ cases (where inconsistencies appear). This helps in understanding how robust an LLM is for a given dataset.
Their experiments with various LLMs (including GPT-3.5-Turbo, Gemma-2-9B, Llama-3.1-8B, Mistral-7B, Qwen2.5-7B, and Deepseek-7B) and eight benchmark datasets revealed significant findings:
- The robustness of LLMs in feature engineering varies greatly across different datasets, suggesting that an LLM’s prior knowledge of a domain plays a crucial role.
- Simply adding more descriptions or examples doesn’t always improve performance; in some cases, it can even degrade it. The quality of examples is far more critical for enhancing robustness.
- The impact of providing a few examples (few-shot learning) can be inconsistent, sometimes even leading to worse results than providing no examples (zero-shot learning). The effectiveness heavily depends on the quality of the samples.
Evaluating Feature Quality and Performance
Beyond diagnosis, the framework also evaluates the quality of the features generated by LLMs. By aligning LLM-generated features with the ‘golden’ variables, relations, and values identified by the framework, the researchers can assess how well these features contribute to predictive performance.
They demonstrated that using high-quality features, identified through their evaluation scheme, can significantly improve the prediction performance of state-of-the-art methods by up to 10.52%. This highlights that by understanding and addressing the robustness issues, LLM-driven feature engineering can become a more reliable and effective tool.
Also Read:
- AI Agents Transform Data Analysis: A Comprehensive Overview
- Structural Reward Models: A New Approach to Interpretable and Efficient AI Evaluation
Future Directions
While promising, the study acknowledges limitations, such as the LLM’s dependence on dataset characteristics and the challenges in fully mitigating risks associated with sample quality. Future research could explore zero-shot feature engineering or incorporate more complex feature representations to further generalize and expand the framework. The work emphasizes the importance of systematically evaluating LLM reliability to unlock their full potential in real-world data science applications.
